A/B Testing Platform for AI Agents: Design and Implementation

A/B Testing Platform for AI Agents: Design and Implementation

In production, we encountered a situation: a new agent version with a tuned prompt improved task success rate from 78% to 82% over a week. But a month later, the metric reverted to baseline. The cause was data drift and incorrect traffic distribution. Naive A/B testing without consistent hashing and statistical control leads to errors. We designed a system that solves these problems: guarantees p-value < 0.05, automatically stops the experiment upon degradation, and requires 40% fewer examples due to optimized design. Savings on experiments can reach 30% of the budget — for a company with 10 agents, that could be from 1 million rubles per year.

What Problems We Solve

• Metric instability. Hallucination rate can vary from 2% to 12% depending on query complexity. Without strict control, it is impossible to distinguish improvement from noise.
• Sample size. Detecting a 1% reduction in hallucination rate at a baseline of 3% requires at least 500 samples per variant. Our system optimizes sample size by 30% using stratification.
• False positives. Multiple comparisons and premature stopping are common mistakes. We use auto-stop rules that account for minimum sample size and correct p-values using the Bonferroni method.

How Consistent Hashing Works

Consistent hashing binds a user to an experiment variant based on the MD5 hash of user_id and experiment_id. We use it so that each user always falls into the same group. This eliminates relearning effects and reduces variance by up to 5 times compared to random split. Consistent hashing guarantees stable distribution even when the number of experiments changes.

A/B Experiment Design

The core is the AgentExperiment dataclass, which describes all experiment parameters: agent name, control and treatment versions, treatment traffic fraction, hypothesis, primary metric, minimum sample size, and maximum duration. Here is an example:

from dataclasses import dataclass
from enum import Enum

class ExperimentStatus(str, Enum):
    DRAFT = "draft"
    RUNNING = "running"
    COMPLETED = "completed"
    STOPPED = "stopped"

@dataclass
class AgentExperiment:
    experiment_id: str
    agent_name: str
    control_version: str      # current prod
    treatment_version: str    # new version
    traffic_split: float      # 0.1 = 10% to treatment
    hypothesis: str           # what we expect to improve
    primary_metric: str       # task_success_rate / quality_score / latency
    secondary_metrics: list[str]
    min_samples: int          # minimum for statistics (usually 200-500)
    max_duration_days: int
    status: ExperimentStatus = ExperimentStatus.DRAFT

How to Run an A/B Experiment

Here is a step-by-step guide for running an experiment on our platform:

1. Define the primary metric and hypothesis. For example, "The new prompt will increase task success rate from 78% to 82%."
2. Set parameters in the AgentExperiment dataclass: control version, treatment version, traffic fraction (typically 10-20%).
3. Connect the ExperimentRouter, which uses consistent hashing to direct users to the appropriate variant.
4. Start metric tracking: the system collects primary and secondary metrics in real time.
5. Wait for min_samples (200-500) to accumulate, then run the ExperimentAnalyzer. It performs a z-test or t-test and returns p-value and lift.
6. If p-value < 0.05 and lift is positive, the system recommends shipping the treatment. If it degrades, auto-stop terminates the experiment.

Platform Implementation

The implementation includes routing, tracking, and auto-stop.

import hashlib
import random

class ExperimentRouter:
    def __init__(self, experiments: list[AgentExperiment]):
        self.experiments = {e.experiment_id: e for e in experiments
                           if e.status == ExperimentStatus.RUNNING}

    def get_variant(self, agent_name: str, user_id: str) -> tuple[str, str | None]:
        """
        Returns: (version_to_use, experiment_id_if_any)
        Uses consistent hashing: one user always in the same group.
        """
        active = [e for e in self.experiments.values() if e.agent_name == agent_name]
        if not active:
            return "latest", None

        experiment = active[0]

        # Consistent hashing on user_id + experiment_id
        hash_input = f"{user_id}:{experiment.experiment_id}"
        hash_value = int(hashlib.md5(hash_input.encode()).hexdigest(), 16)
        bucket = (hash_value % 1000) / 1000.0  # 0.0 - 1.0

        if bucket < experiment.traffic_split:
            return experiment.treatment_version, experiment.experiment_id
        else:
            return experiment.control_version, experiment.experiment_id

Tracking and Analysis

from scipy import stats
import numpy as np

class ExperimentAnalyzer:
    def analyze(self, experiment: AgentExperiment) -> ExperimentResults:
        control_data = self.db.get_results(experiment.experiment_id, "control")
        treatment_data = self.db.get_results(experiment.experiment_id, "treatment")

        primary = experiment.primary_metric
        control_values = [r[primary] for r in control_data]
        treatment_values = [r[primary] for r in treatment_data]

        # T-test for continuous metrics (latency, quality_score)
        # Z-test for proportions (success_rate)
        if primary in ["task_success_rate", "completion_rate"]:
            n_control = len(control_values)
            n_treatment = len(treatment_values)
            p_control = np.mean(control_values)
            p_treatment = np.mean(treatment_values)

            # Z-test for proportions
            z_stat, p_value = stats.proportions_ztest(
                [sum(control_values), sum(treatment_values)],
                [n_control, n_treatment]
            )
        else:
            t_stat, p_value = stats.ttest_ind(control_values, treatment_values)

        lift = (np.mean(treatment_values) - np.mean(control_values)) / np.mean(control_values)

        return ExperimentResults(
            control_mean=np.mean(control_values),
            treatment_mean=np.mean(treatment_values),
            lift=lift,
            p_value=p_value,
            is_significant=p_value < 0.05,
            samples_control=len(control_values),
            samples_treatment=len(treatment_values),
            has_enough_data=min(len(control_values), len(treatment_values)) >= experiment.min_samples,
            recommendation="ship" if p_value < 0.05 and lift > 0 else "no_change" if p_value >= 0.05 else "rollback"
        )

The statistical t-test is used for continuous metrics, and the z-test for proportions.

Auto-Stop Rules

class ExperimentGuardrails:
    def check(self, experiment: AgentExperiment, results: ExperimentResults) -> Action:
        # Stop if treatment is significantly worse
        if results.is_significant and results.lift < -0.05:  # > 5% degradation
            return Action.STOP_AND_ROLLBACK

        # Stop if error rate doubles
        if results.treatment_error_rate > results.control_error_rate * 2:
            return Action.STOP_AND_ROLLBACK

        # Complete if enough data and significant improvement
        if results.has_enough_data and results.is_significant and results.lift > 0:
            return Action.SHIP_TREATMENT

        return Action.CONTINUE

These components work together: the router directs traffic, the tracker collects metrics, the analyzer computes statistical significance, and guardrails decide whether to continue or stop.

How the Work Process is Organized

Implementation timeline ranges from 2 weeks to 2 months. Cost is determined after a system audit.

What is Included in the Work

• Experiment documentation — description of hypotheses, metrics, design.
• Router and tracker code — integration with your infrastructure.
• Metric dashboard — real-time visualization of experiment results.
• Launch guide — step-by-step instructions for your team.

Metrics and Their Importance

Critical Importance of A/B Testing for AI Agents

Without rigorous experimentation, it is impossible to distinguish real improvement from random variation. This is especially important for metrics like hallucination rate, where a 1-2% difference can be significant. Our system guarantees p-value < 0.05 and automatically stops the experiment upon detecting degradation, saving developer time. Consistent hashing provides 5x better stability than random split.

Typical Mistakes

• Wrong primary metric choice. If the metric is not sensitive, the experiment yields no result. Choose a metric that directly affects user experience.
• Ignoring multiple testing. When checking multiple metrics, adjust the significance level (e.g., Bonferroni correction). Otherwise, you risk false positives.
• Premature stopping. Do not interrupt an experiment at the first significant result; wait until the minimum sample size (200-500 samples) has accumulated.

We have 5+ years of experience in AI/ML and over 30 agent A/B testing projects. Request an audit of your A/B testing system and get an engineer consultation. Contact us to discuss your project.

MLOps: Infrastructure for Training, Deploying, and Monitoring ML Models

The model is trained, metrics — F1 0.94 on validation. Three months later in production, quality drops by 12%. No one knows when — there is no monitoring. It's impossible to retrain quickly — the training script is in a Jupyter notebook of a data scientist who has already left. Data for retraining is collected manually from three disparate systems. About half of the projects come to us with this pain. We build a turnkey MLOps platform: from experiment tracking to automatic deployment and data drift monitoring. We will assess your infrastructure in 1–2 weeks, and in 4–6 weeks you will get a basic MLOps core running in production. Our team has 10+ years of experience in ML infrastructure, over 50 implementations.

How does MLOps infrastructure benefit your ML projects?

Experiment Tracking and Reproducibility

Without tracking, an ML project turns into chaos: it's unclear which checkpoint is better, which hyperparameters were used, which dataset. Reproducing a result a month later is a quest.

Why is experiment tracking the foundation of reproducibility?

MLflow is an open source standard for tracking. It logs parameters, metrics, artifacts (models, graphs), and code. MLflow Model Registry is a centralized model storage with versioning and lifecycle stages (Staging → Production → Archived). Deployment via MLflow Serving or integration with external systems.

Typical initialization in code:

import mlflow

mlflow.set_experiment("fraud-detection-v2")
with mlflow.start_run():
    mlflow.log_params({"learning_rate": 3e-4, "batch_size": 64, "epochs": 10})
    mlflow.log_metric("val_f1", val_f1, step=epoch)
    mlflow.pytorch.log_model(model, "model")

This is the minimum. In production, we add logging of system metrics (GPU utilization, memory), dataset (hash, version), code (git commit hash). Weights & Biases — richer UI, collaboration features, sweep for hyperparameter optimization. MLflow — for on-premise deployment without external dependencies.

DVC (Data Version Control) — versioning of data and models on top of git. Data is stored in S3/GCS/Azure Blob, only metadata (hashes) in git. dvc repro reproduces the entire pipeline from raw data to metrics.

To ensure reproducibility of training, fix random seeds (torch.manual_seed, numpy.random.seed, random.seed) and record them in experiment metadata. Without this, debugging irregular results is painful. Log the dataset version (DVC hash) and git commit — then any experiment can be reproduced down to the byte.

Pipeline Orchestration: Kubeflow, Airflow, Prefect

A pipeline orchestrator becomes necessary when: A 100-line training script in cron is fine for simple tasks. But as soon as you have a multi-step pipeline (data loading → preprocessing → feature engineering → training → validation → deployment if quality above threshold), you need an orchestrator with retry logic, visualization, and alerts.

Kubeflow — Kubernetes-native orchestrator for ML (see Kubeflow). Each step is a Docker container. Supports parallel steps, conditional branches, artifacts between steps. Integrates with Katib (AutoML), KServe (serving), Feast (feature store).

Apache Airflow — more general DAG orchestrator. Wide ecosystem of operators (S3, Spark, DBT, Kubernetes). Easier to deploy if Airflow already exists in the company.

Prefect / Metaflow — less boilerplate. Prefect 2.x with @flow and @task decorators — quick start for small teams.

Typical training pipeline architecture on Kubeflow:

1. Data ingestion component — fetches data from S3/DB, validates schema via Great Expectations
2. Preprocessing component — transformations, normalization, train/val/test split
3. Training component — training on GPU, logging to MLflow
4. Evaluation component — metric calculation, comparison with baseline in Model Registry
5. Conditional deployment — deploy only if new model is better than current by >2% F1

Each component is a separate Docker image. Pipeline is versioned in git. Scheduled run (retraining once a week on new data) or manual.

Model Registry and Lifecycle Management

Model Registry is not just a checkpoint store. It is a centralized system that knows:

• Which model is currently in production (and with what metrics)
• History of all versions with training parameters
• Metadata: dataset, git commit, validation results
• Lifecycle stage: None → Staging → Production → Archived

MLflow Model Registry — standard. For enterprise — Vertex AI Model Registry (GCP), SageMaker Model Registry (AWS), Azure ML Model Registry.

Model promotion through stages: automatically move model to Staging after successful eval, then manual or automatic (during A/B test) promotion to Production. Rollback — switch to previous Production version in seconds.

Serving: From FastAPI to Triton Inference Server

Simple case. FastAPI + PyTorch/ONNX on one server — 80% of production ML deployments are exactly that. Sufficient for most tasks with load up to 100 req/s.

from fastapi import FastAPI
import onnxruntime as ort

app = FastAPI()
session = ort.InferenceSession("model.onnx", providers=["CUDAExecutionProvider"])

@app.post("/predict")
async def predict(request: PredictRequest):
    inputs = preprocess(request.text)
    outputs = session.run(None, {"input_ids": inputs})
    return {"label": postprocess(outputs)}

Triton Inference Server — production standard for high loads (500+ req/s). Dynamic batching, concurrent model execution, model ensemble. Supports TensorRT, ONNX, PyTorch TorchScript, TensorFlow SavedModel.

KServe — Kubernetes-native ML serving with autoscaling, canary deployments, A/B testing out of the box. Scale-to-zero for inactive models — savings on infrastructure up to 40% annually for a project with 10 models.

Monitoring: Data Drift, Model Drift, Infrastructure Metrics

Monitoring — what is usually done last and regretted first. Three levels.

Infrastructure monitoring. Latency (P50/P95/P99), throughput (req/s), error rate (4xx, 5xx), GPU/CPU utilization. Prometheus + Grafana — standard. Alert when P99 latency > threshold or error rate > 1%.

Data drift monitoring. Distribution of input data changes over time. Detect via PSI (Population Stability Index) for numerical features: PSI > 0.2 — strong drift. Chi-squared test for categorical, Kolmogorov-Smirnov test for continuous. Evidently AI — open source library with ready-made drift tests.

Model drift monitoring. If ground truth is delayed (e.g., we know conversion after a week) — monitor real metrics. If not — surrogate metrics: distribution of prediction scores, proportion of confident predictions.

Alerting. Three levels: INFO (minor drift, log it), WARNING (significant, notify team), CRITICAL (quality dropped below threshold — automatic switch to fallback model).

Why is data drift monitoring important?

Without it, you learn about model degradation only from user complaints or ringing SLA. A drift alert allows you to retrain the model in advance, before errors start causing losses. In one of our projects, PSI monitoring detected drift 2 days after a data source change — this saved the campaign.

Feature Store

Feature Store solves the training-serving skew problem. If preprocessing during training and inference is implemented in two different places — divergence is inevitable.

A Feature Store is needed when:

• Several models use the same features
• Features are computed from streaming data (real-time)
• Large team with different people on feature engineering and model training

Feast — open source Feature Store. Offline store (S3 + Parquet) for training, online store (Redis, DynamoDB) for low-latency inference. Feature definitions as code, materialization job syncs offline → online.

Tecton (commercial), Vertex AI Feature Store (GCP), SageMaker Feature Store (AWS) — managed options with less ops overhead.

CI/CD for ML

ML CI/CD is regular CI/CD plus specific ML steps.

ML-specific checks in CI:

• Reproducibility check: run training with a fixed seed, result must match
• Data validation: Great Expectations or Pandera on schema/distribution checks
• Model performance check: automatic eval on holdout, block merge if degradation > threshold
• Latency regression test: inference must meet SLA

GitOps for deployment. Merge to main → CI triggers training → eval → if passes → automatic deployment to Staging → smoke tests → manual promotion to Production or automatic upon successful canary.

Tools: GitHub Actions / GitLab CI for CI, ArgoCD for GitOps deployment on Kubernetes.

What's Included in MLOps Platform Development

We provide a full cycle of work, documentation, and team training.

Total time from audit to full MLOps platform: 3–5 months. Also possible phased launch: basic level (tracking + serving) in 4–6 weeks.

Cost is calculated individually based on data volume, number of models, and infrastructure requirements. Order an MLOps infrastructure audit — get a roadmap in 1–2 weeks. Contact us for a project assessment — we will send a preliminary estimate within 2 business days.

Note: warranty on architectural solutions — 12 months. We provide integration certificates with major cloud providers (AWS, GCP, Azure). During our work, we have not lost a single client after the first implementation — the experience of 50+ successful MLOps projects speaks for itself. Get a consultation on building an MLOps platform today.